ul 2004 
omof 2 
of 0.36 
se by a 
he other 
“ast and 
ficantly 
he HRS 
cost of 
overall 
eters Is 
ne to an 
tigated. 
s, point 
med to 
stics of 
n RMS 
superior 
hile the 
ch i.e. 
se to the 
  
ply the 
rmance 
results. 
ogy as 
eese test 
st areas 
latching 
rom the 
levation 
pproach 
pproach 
‘able 4, 
ing was 
ised by 
pixels 1s 
y major 
  
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004 
  
v 
AN Li 
  
  
  
  
Figure 1: Detailed test areas showing hilly (left), mountainous (mid) and urban (right) terrain in anaglyph presentation 
(red: forward image, green: backward image). 
While for the rural and the urban area only 1,45% and 2,83% of 
the entire pixels were not matched, a failure rate of 10,82% was 
achieved for the mountainous test arca. This is due to the large 
parallaxes (dissimilarities) between the 2 stereo images, as they 
are present in case of mountainous terrain and large base-to- 
height ratio of the data. 
  
Rural area 1,45 % 
Mountainous area 10,82 % 
Urban area 2,83 % 
Table 4: Matching performance for HRS stereo data. 
  
  
  
  
  
  
The quality of the surface models resulting from these stereo 
matching results is checked through comparison with the 
reference elevation model, i.e. through calculation of height 
differences. However, the reference elevation model does not 
represent the a surface model, but rather a ground model, 
excluding objects like trees or buildings. 
The digital surface models resulting for the selected test areas 
as well as the difference elevation models, which were 
determined with respect to the given reference DEM, are shown 
in Figure 2. Moreover, the overall statistics of elevation 
differences are summarized in Table 5. The following 
comments can be made: 
Hilly test area: The elevation errors clearly correlate with 
ground cover, i.e. positive elevation differences in the order 
of more than 5 meters are frequently achieved for forested 
parts of the test area. This also results in a bias in the height 
differences of 1.6 meters, while the standard deviation is 
5.5 meters. 
Mountainous test area: For this area large height errors are 
party achieved over the central mountainous area, which 
covers an elevation range between 76 and 1206 meters. 
Local areas with rather extreme elevation errors of up to 
300 meters distinctly degrade the standard deviation to 35.8 
meters only for this test area. 
Urban area: For the built-up areas height differences of more 
  
the buildings in this area. This leads to a bias of 9.3 meters, 
while the standard deviation is 8 meters. However, 
individual buildings can not really be discriminated 
  
Model Area Mean | Std.Dev.| Min. Max. 
  
  
  
  
  
  
  
  
  
2 1.6 5.5 -78.2 48.0 
HRS1-HRS2 4 0.3 35.8 -300.4 224.5 
6 9.3 8.0 -23.1 46.3 
  
  
than 5 meters are widely achieved, reflecting the heights of 
Table 5: Summary of elevation difference statistics for 
investigated test cases. 
A visual quality check can be made through stereo ortho 
photos, which are generated from the input stereo images and 
using these surface models. Location differences of these ortho 
photos indicate elevation errors in the surface model. A 
superposition of the HRS ortho photos in red (forward image) 
and green (backward image) is shown in Figure 3 for the rural 
and the mountainous test area, respectively. A satisfactory 
correspondence is achieved for the rural area, while extreme 
differences result for the mountainous area. 
4.0. DSM from THR-HRS stereo pair 
The multi-sensor stereo model comprising the THR image in 
conjunction with the HRSI stereo image was used in order to 
investigate the benefit of the THR supermode product with 
respect to DSM generation. For image matching, the HRSI 
stereo image was over-sampled and coarsely registered to the 
geometry of the THR image. 
This approach preserves the high resolution of the THR image. 
However, a distinctly different level of detail is inherent to the 
resulting stereo pair. This is shown in Figure 4, where sub- 
windows of the stereo images are presented for a built-up area 
and a rural area. It is obvious, that many details being visible in 
the high-resolution THR image disappear in the over-sampled 
HRS1 image. Successful matching cannot be expected for such 
features.